Xiaomi Unveils Xring O3, O100, and D100: Three In-House AI Chips Build the Computing Foundation for Its Human-Car-Home Ecosystem

After more than five years of sustained work in proprietary silicon, Xiaomi has unveiled a new generation of its Xuanjie chip lineup: the Xuanjie O3, Xuanjie O100, and Xuanjie D100. The three processors target different computing needs across mobile devices, large-model acceleration, and intelligent driving, and Xiaomi says all of them have already completed silicon validation.

The launch marks a broader shift in Xiaomi’s in-house chip roadmap. What began as efforts around smartphone SoCs and communication chips is now expanding into a wider AI computing framework designed to support the company’s “Human x Car x Home” ecosystem.

According to Xiaomi, the company has spent more than five years since restarting large-chip development, with cumulative R&D investment exceeding RMB 21 billion and a chip engineering team approaching 3,000 people. Earlier products, including the Xuanjie O1—described as the first 3nm advanced-process SoC from mainland China—and the Xuanjie T1 communication chip, helped establish the technology base and talent pool behind this latest release.

Xuanjie O3: Xiaomi’s first flagship AI SoC

The centerpiece of the announcement is the Xuanjie O3, positioned as Xiaomi’s first flagship-grade AI SoC for premium mobile devices. Built on a 3nm process, the chip measures 133mm² and integrates 24 billion transistors, a 26% increase over the earlier Xuanjie O1.

Xiaomi says the O3 improves transistor density by 5% through the use of Top Channelless channel-elimination technology together with a self-developed standard cell library. On the CPU side, it features a ten-core all-big-core design with a peak clock speed of 4.35GHz. The company also cited an AnTuTu maximum benchmark score of 5.22 million.

For graphics, the chip debuts a 16-core G2-Ultra NX GPU. Xiaomi claims graphics performance is close to doubling compared with the previous generation, while GPU power consumption is reduced by 64%.

Memory is another major focus. The Xuanjie O3 is said to be the first in the industry to support LPDDR6 memory, raising bandwidth to 113.8GB/s and cutting memory access latency to 82ns.

Its AI subsystem has also been substantially redesigned. Xiaomi says the chip’s NPU architecture has been deeply reworked, bringing Tensor computing performance to 200 TOPS, alongside a clear expansion in vector computing resources. Combined with Xiaomi’s MiMo large-model customized 5-value quantization scheme and hardware-based Huffman lossless compression, on-device AI inference speed improves by 45%, while power use during inference drops by 26%.

Xiaomi also emphasized that AI acceleration is not limited to the NPU alone. The CPU, GPU, imaging ISP, display engine, and audio DSP all include dedicated AI acceleration hardware, enabling what the company describes as end-to-end AI enhancement across the entire chip pipeline.

• Process: 3nm

• Chip area: 133mm²

• Transistor count: 24 billion

• CPU: 10-core all-big-core design, up to 4.35GHz

• GPU: 16-core G2-Ultra NX

• Memory: LPDDR6, 113.8GB/s bandwidth

• AI performance: 200 TOPS Tensor compute

In imaging, the O3 integrates Xiaomi’s fifth-generation in-house ISP, with support for up to 432-megapixel capture and 4K 60fps AI night video. On the security side, it includes a self-developed RISC-V security core and has received both Chinese commercial cryptography certification and CCRC EAL5+ certification.

Xiaomi says the Xuanjie O3 will debut in its upcoming flagship foldable, the Xiaomi 18 Fold.

Xuanjie O100: Built to accelerate on-device large-model inference

Alongside the O3, Xiaomi introduced the Xuanjie O100, an AI accelerator focused on the bandwidth demands of local large-model inference. The company describes it as the world’s first 6nm 3D wafer-level stacked AI accelerator chip.

The O100 uses a combination of Wafer-on-Wafer packaging, Hybrid Bonding, and a face-to-face metal interconnect architecture. Xiaomi’s goal here is to tackle the “memory wall” problem through a near-memory computing approach.

According to the company, the chip delivers up to 1.22TB/s of bandwidth—16 times the memory bandwidth of mainstream flagship smartphones. It also includes a 14-core high-performance NPU and a high-bandwidth matrix bus capable of dynamic topology switching.

When paired with the Xuanjie O3, the O100 can push on-device large-model inference to as much as 330 tokens per second. Xiaomi says commercial deployment is planned for next year.

• Process: 6nm 3D wafer-level stacking

• Bandwidth: 1.22TB/s

• NPU: 14-core high-performance NPU

• On-device inference speed: up to 330 tokens/s

Xuanjie D100: Targeting high-compute intelligent driving

For the automotive segment, Xiaomi announced the Xuanjie D100, an AI chip aimed at intelligent driving systems. It is positioned as China’s first 3nm AI chip for smart driving applications.

The D100 combines a 20-core high-performance CPU with a 16-core high-compute NPU. It supports up to 160GB of unified memory and is designed to locally deploy large models with up to 200 billion parameters.

Xiaomi plans to bring the D100 into commercial use next year, filling an important gap in its self-developed compute strategy for smart vehicles.

• Process: 3nm

• CPU: 20-core high-performance CPU

• NPU: 16-core high-compute NPU

• Unified memory: up to 160GB

• Model deployment: supports local deployment of 200 billion-parameter models

Xiaomi’s chip strategy moves beyond devices into ecosystem AI infrastructure

Xiaomi founder Lei Jun said the Xuanjie O3, O100, and D100 together span mobile terminals, smart vehicles, and smart home scenarios, forming the AI computing foundation for Xiaomi’s full “Human x Car x Home” ecosystem.

That framing suggests the Xuanjie program is no longer just a hardware self-sufficiency effort. It is increasingly being positioned as the physical computing layer underpinning Xiaomi’s broader AI ambitions.

As these chips move toward commercial deployment, Xiaomi is taking a more concrete step toward building end-side AI capabilities across all of its major product categories.